I build AI that earns developer trust.

// now · the through line

Three questions drive the work.

One thread runs through everything I have built: how do we make AI assistance in software engineering something a developer can rely on, and prove that it is.

// 2025 · jetbrains research

Research at production scale.

As a research intern at JetBrains, I asked a direct question: can a model's confidence tell a developer when to trust its code. The answer was not simple. Post-hoc calibration sharpens the signal, but a model's confidence is not the same thing as a developer's acceptance.

// 2025 · a bet placed early

When agents multiplied, I asked what breaks first.

My answer was safety. AgentGuard, my first single-author paper, argued for verifying agents at runtime rather than trusting them by default. It builds a probabilistic model of an agent's behavior from its execution traces and checks quantitative properties against it. That question grew into a research line: an automated abstraction mechanism (TriCEGAR) and a formal theory of what an agent can be once you constrain its memory.

// 2025 · the long one

If the ruler is wrong, every measurement is.

A review of 273 benchmarks for AI in software engineering, a semantic search tool to navigate them (BenchScout), and a protocol for repairing the ones the field still relies on (BenchFrame). When we rebuilt HumanEval with correct solutions and real edge cases, the leaderboard collapsed.

// 2024 · learning the craft

Where the research habit formed.

I joined the AISE Lab under Prof. Maliheh Izadi and started where curiosity pointed: hyperdimensional computing as a lightweight way to model how developers behave inside an IDE. Two early papers, a lot of freedom. My first paper, on making automatically generated unit tests understandable, landed at ICSE 2025. The idea was simple: a test only helps if a person can read it.

// generated by a search-based tool
void test0() { ... }

// after UTGen: named for what it verifies
void testEqualsWithDifferentMinDamageValues() {
  // given / when / then, in plain terms
  Weapon sword = new Weapon("sword", 12);
  ...
}

In a controlled study, developers fixed up to 33% more bugs and spent up to 20% less time with the enhanced tests.

// 2023 · the groundwork

A degree, and a question about memory.

A BSc in Computer Science and Engineering at TU Delft (2023–2026), with a mathematics minor at the University of Amsterdam for the theory I kept reaching for. My bachelor thesis asked whether you can tell if a code model was trained on a given file. You often can, and the signal survives even after the code is refactored. The trick was to perturb code along its syntax tree instead of its raw tokens, so the calibration samples stay valid.

// 2022 · where it started

Structure out of noise.

It began at Syntho, first as a high-school intern and then as a software engineer, working on synthetic data and fairness. Designing methods to generate data that is useful without being harmful is where building and researching stopped being two separate activities for me. Everything after this is a variation on the same instinct: understand a system well enough to make it better.

// away from the screen

The rest of it.

I sing and play music, I train and dance bachata, and I cook. I move between three languages depending on the room. The through line is the same as the work: I like understanding how something functions well enough to improve it.